Leakage-Conscious Machine Learning and MCDM Analysis of Y/Bi-Substituted Bi-2212 Ceramics

Abstract

Y/Bi substitution in Bi-2212 ceramics produces a peaked mechanical response within a narrow composition window. This materials informatics study applies a compact, leakage-conscious machine-learning protocol to the published indentation dataset to recover and explain that compositional trend. The modelling table comprises 35 load-level rows drawn from seven ceramic compositions across five Vickers loads, with sample-level EDX descriptors assigned to the five load rows of each composition; validation followed leave-one-sample-out splitting at the composition level. Five descriptors were used: x, F, xF, Cu/Bi and Y/Bi. Ridge regression and Extra Trees regression were evaluated with fixed, a priori parameters. For Vickers hardness, Extra Trees gave MAE = 0.0155 GPa, RMSE = 0.0189 GPa and R² = 0.8923 under composition-held-out validation. The same compact protocol recovered the measured trend for stiffness-related descriptors; fracture toughness proved harder to capture. TreeSHAP analysis of the Extra Trees hardness model identified substitution-related descriptors as the dominant explanatory variables: the EDX-derived Y/Bi ratio and nominal x carried the largest mean contributions, followed by Cu/Bi, the load-coupled xF term and the applied load. A TOPSIS ranking of the seven compositions over the five indentation-derived descriptors placed Y-2 first, consistent with both the measured trend and the SHAP-based explanation. The contribution is methodological: a leakage-conscious validation protocol, a compact and auditable descriptor set, an explainable-AI interpretation of the fitted model, and an independent decision-analysis check are added to the original experimental measurements.